Kuis Decoding Lookahead

Jawaban dan penjelasan muncul saat Anda mengikuti kuis.

1. Apa yang membedakan decoding lookahead dari decoding spekulatif standar?
   • Hal ini memerlukan dua model yang dilatih secara terpisah
   • Ia tidak menggunakan rancangan model terpisah — hanya model target itu sendiri
   • Ini hanya berfungsi untuk tugas terjemahan
   • Ini mengkuantisasi model secara permanen
2. Metode numerik manakah yang mendukung decoding lookahead?
   • Penurunan gradien
   • Transformasi Fourier
   • Pengelompokan K-means
   • Iterasi Jacobi (titik tetap).
3. Apa dua cabang paralel yang berjalan pada setiap langkah?
   • Cabang pelatihan dan cabang pengujian
   • Encoder dan decoder
   • Cabang lookahead dan cabang verifikasi
   • Model rancangan dan model target
4. Apa yang disimpan di 'n-gram pool'?
   • Urutan multi-token yang masuk akal terlihat selama iterasi paralel
   • Riwayat obrolan pengguna
   • Bobot model
   • Masker perhatian
5. What supports correctness when lookahead decoding commits a guessed n-gram prefix?
   • Using the same model makes every unverified guess correct
   • Every candidate in the n-gram pool must be committed
   • Verification against the target's greedy predictions or an appropriate distribution-preserving sampling procedure
   • A larger window proves the generated statements are true
6. What resource tradeoff is important when deploying lookahead decoding on already busy hardware?
   • Candidate generation uses no compute or memory
   • Additional parallel work may offset the benefit of fewer sequential passes
   • Every larger window must lower wall-clock latency
   • The target weights are automatically compressed to zero size
7. Why should a lookahead candidate remain provisional until target verification?
   • Parallel iteration guesses can differ from the target's valid continuation
   • Every cached n-gram has already been proven correct by storage alone
   • The lookahead branch permanently changes the target weights
   • Verification is needed only to determine the output filename
8. Does a standard training-free lookahead iteration update the target model's learned weights?
   • Yes; each candidate applies a gradient update to every parameter
   • Yes; verified n-grams become a new training dataset immediately
   • Only the tokenizer's vocabulary is retrained at every step
   • No; it changes candidate token states and verifies outputs while keeping the model weights fixed